Gazpacho and summer rash: lexical relationships from temporal patterns of web search queries

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1046–1055, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Gazpacho and summer rash: lexical relationships from temporal patterns of web search queries Enrique Alfonseca Massimiliano Ciaramita Keith Hall Google Z¨urich, Switzerland ealfonseca@google.com, massi@google.com, kbhall@google.com Abstract In this paper we investigate temporal pat- terns of web search queries. We carry out several evaluations to analyze the proper- ties of temporal profiles of queries, reveal- ing promising semantic and pragmatic re- lationships between words. We focus on two applications: query suggestion and query categorization. The former shows a potential for time-series similarity mea- sures to identify specific semantic relat- edness between words, which results in state-of-the-art performance in query sug- gestion while providing complementary information to more traditional distribu- tional similarity measures. The query cat- egorization evaluation suggests that the temporal profile alone is not a strong in- dicator of broad topical categories. 1 Introduction The temporal patterns of word occurrences in hu- man communication carry an implicit measure of their relationship to real-world events and behav- ioral patterns. For example, when there is an event affecting a given entity (such as a natural disaster in a country), the entity name will turn up more frequently in human conversation, newswire arti- cles and web documents; and people will search for it more often. Two entities that are closely related in the real world, such as the name of a country and a prominent region inside the coun- try are likely to share common events and there- fore be closely associated in human communica- tion. Finally, two instances of the same class are also likely to share common usage patterns. For example, names of airlines or retail stores are more likely to be used by day rather than by night (Chien, 2005). In this paper we explore the linguistic relation- ship between phrases that are judged to be sim- ilar based on their frequency time series correla- tion in search query logs. For every phrase1 avail- able in WordNet 3.02 (Miller, 1995), we have ob- tained its temporal signature from query logs, and calculated all their pairwise correlations. Next, we study the relationship in the top-ranked pairs with respect to their distribution in WordNet and a human-annotated labelling. We also discuss possible applications of this data to solve open problems and present the results of two experiments: one where time series corre- lations turned out to be highly discriminative; and another where they were not particularly informa- tive but shed some light on the nature of temporal semantics and topical categorization: • Query suggestion, i.e. given a query, generate a ranked list of alternative queries in which the user may be interested. • Query categorization, i.e. given a predefined set of categories, find the top categories to which the query can be assigned. Finally, we illustrate with an example another ap- plication of time series in solving information ex- traction problems. Although query logs are typically proprietary data, there are …

Content truncated to 500 words